Interpretable Dynamic Brain Network Analysis With Functional and Structural Priors.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Medical Image Analysis

Background:

• Dynamic functional brain networks (DFBNs) are crucial for understanding brain activity and detecting disorders.
• Current DFBN analysis methods often lack interpretability due to reliance on data-driven models and neglect of prior brain knowledge.
• Extracting complex spatio-temporal features from DFBNs remains challenging.

Purpose of the Study:

• To propose an interpretable spatio-temporal tensor graph convolutional network for DFBN analysis.
• To enhance the interpretability and diagnostic performance of DFBN analysis for brain disorders.
• To effectively capture and extract spatio-temporal topological features from DFBNs.

Main Methods:

• Incorporated functional and structural priors to construct a hierarchical DFBN representation with brain region clustering.
• Developed a tensor graph convolutional network with intra-graph and inter-graph propagation for spatio-temporal feature extraction.
• Utilized a functional subnetwork constraint and self-attention for feature enhancement and fusion.

Main Results:

• The proposed method achieved competitive diagnostic performance on epilepsy, ADNI, and ABIDE datasets.
• Demonstrated network-level interpretability for brain disease diagnosis.
• Effectively captured spatio-temporal topology and enhanced feature consistency within subnetworks.

Conclusions:

• The interpretable spatio-temporal tensor graph convolutional network offers a promising approach for DFBN analysis in brain disorder detection.
• Integrating prior knowledge significantly improves model interpretability and diagnostic accuracy.
• The method provides valuable insights into network-level alterations associated with brain diseases.